From slurs to slots: LLM masking and Telephone-Game Augmentation for multiclass hate speech detection

We present an offline, online hybrid for multiclass hate speech detection. Large language models (LLMs) are used only at training time to mask volatile slurs and targets into normalized placeholders (e.g., [SLUR:] , [TARGET:] ) and to generate label-preserving paraphrases through a Telephone-Game augmentation (TGA). In inference, lightweight classifiers operate on masked inputs with predictable latency. On a strictly held-out test set, masking improves binary macro-F1 by ＋0.073 and accuracy by ＋0.071 on average in five small models. In multiclass, masking yields model-dependent gains; for example, Linear SVM improves by ＋0.141 macro-F1, while TGA further boosts minority and low-regime performance where only about 20 samples per class are available. Under equal budgets, TGA (RAG) outperforms a back-translation baseline by ＋0.109 macro-F1. We release a lexicon, a placeholder taxonomy, code, and prompts. Results indicate a practical path to robust moderation, showing that LLM cost can be amortized offline, lightweight models can be deployed with stable latency, and the approach can be expanded to new hate categories in a changing environment with limited labeled data.